Alibaba's now famed Qwen AI improvement staff has carried out it once more: somewhat greater than a day in the past, they launched the Qwen3.5 Medium Mannequin sequence consisting of 4 new massive language fashions (LLMs) with assist for agentic software calling, three of which can be found for business utilization by enterprises and indie builders underneath the usual open supply Apache 2.0 license:
Qwen3.5-35B-A3B
Qwen3.5-122B-A10B
Qwen3.5-27B
Builders can obtain them now on Hugging Face and ModelScope. A fourth mannequin, Qwen3.5-Flash, seems to be proprietary and solely accessible by the Alibaba Cloud Mannequin Studio API, however nonetheless presents a powerful benefit in price in comparison with different fashions within the West (see pricing comparability desk beneath).
However the large twist with the open supply fashions is that they provide comparably excessive efficiency on third-party benchmark assessments to similarly-sized proprietary fashions from main U.S. startups like OpenAI or Anthropic, truly beating OpenAI's GPT-5-mini and Anthropic's Claude Sonnet 4.5 — the latter mannequin which was simply launched 5 months in the past.
And, the Qwen staff says it has engineered these fashions to stay extremely correct even when "quantized," a course of that reduces their footprint additional by decreasing the numbers by which the mannequin's settings are saved from many values to far fewer.
Crucially, this launch brings "frontier-level" context home windows to the desktop PC. The flagship Qwen3.5-35B-A3B can now exceed a 1 million token context size on consumer-grade GPUs with 32GB of VRAM. Whereas not one thing everybody has entry to, that is far much less compute than many different comparably-performant choices.
This leap is made doable by near-lossless accuracy underneath 4-bit weight and KV cache quantization, permitting builders to course of huge datasets with out server-grade infrastructure.
Expertise: Delta power
On the coronary heart of Qwen 3.5's efficiency is a classy hybrid structure. Whereas many fashions rely solely on normal Transformer blocks, Qwen 3.5 integrates Gated Delta Networks mixed with a sparse Combination-of-Consultants (MoE) system.The technical specs for the Qwen3.5-35B-A3B reveal a extremely environment friendly design:
Parameter Effectivity: Whereas the mannequin homes 35 billion parameters in whole, it solely prompts 3 billion for any given token.
Professional Variety: The MoE layer makes use of 256 specialists, with 8 routed specialists and 1 shared skilled serving to to take care of efficiency whereas slashing inference latency.
Close to-Lossless Quantization: The sequence maintains excessive accuracy even when compressed to 4-bit weights, considerably decreasing the reminiscence footprint for native deployment.
Base Mannequin Launch: In a transfer to assist the analysis group, Alibaba has open-sourced the Qwen3.5-35B-A3B-Base mannequin alongside the instruct-tuned variations.
Product: Intelligence that 'thinks' first
Qwen 3.5 introduces a local "Pondering Mode" as its default state. Earlier than offering a ultimate reply, the mannequin generates an inner reasoning chain—delimited by <suppose> tags—to work by advanced logic.The product lineup is tailor-made for various {hardware} environments:
Qwen3.5-27B: Optimized for prime effectivity, supporting a context size of over 800K tokens.
Qwen3.5-Flash: The production-grade hosted model, that includes a default 1 million token context size and built-in official instruments.
Qwen3.5-122B-A10B: Designed for server-grade GPUs (80GB VRAM), this mannequin helps 1M+ context lengths whereas narrowing the hole with the world's largest frontier fashions.
Benchmark outcomes validate this architectural shift. The 35B-A3B mannequin notably surpasses a lot bigger predecessors, reminiscent of Qwen3-235B, in addition to the aforementioned proprietary GPT-5 mini and Sonnet 4.5 in classes together with information (MMMLU) and visible reasoning (MMMU-Professional).
Pricing and API integration
For these not internet hosting their very own weights, Alibaba Cloud Mannequin Studio supplies a aggressive API for Qwen3.5-Flash.
Enter: $0.1 per 1M tokens
Output: $0.4 per 1M tokens
Cache Creation: $0.125 per 1M tokens
Cache Learn: $0.01 per 1M tokens
The API additionally contains a granular Software Calling pricing mannequin, with Net Search at $10 per 1,000 calls and Code Interpreter presently provided for a restricted time for gratis.
This makes Qwen3.5-Flash among the many most reasonably priced to run over API amongst all the main LLMs on the earth. See a desk evaluating them beneath:
Mannequin | Enter | Output | Whole Value | Supply |
Qwen 3 Turbo | $0.05 | $0.20 | $0.25 | |
Qwen3.5-Flash | $0.10 | $0.40 | $0.50 | |
deepseek-chat (V3.2-Exp) | $0.28 | $0.42 | $0.70 | |
deepseek-reasoner (V3.2-Exp) | $0.28 | $0.42 | $0.70 | |
Grok 4.1 Quick (reasoning) | $0.20 | $0.50 | $0.70 | |
Grok 4.1 Quick (non-reasoning) | $0.20 | $0.50 | $0.70 | |
MiniMax M2.5 | $0.15 | $1.20 | $1.35 | |
MiniMax M2.5-Lightning | $0.30 | $2.40 | $2.70 | |
Gemini 3 Flash Preview | $0.50 | $3.00 | $3.50 | |
Kimi-k2.5 | $0.60 | $3.00 | $3.60 | |
GLM-5 | $1.00 | $3.20 | $4.20 | |
ERNIE 5.0 | $0.85 | $3.40 | $4.25 | |
Claude Haiku 4.5 | $1.00 | $5.00 | $6.00 | |
Qwen3-Max (2026-01-23) | $1.20 | $6.00 | $7.20 | |
Gemini 3 Professional (≤200K) | $2.00 | $12.00 | $14.00 | |
GPT-5.2 | $1.75 | $14.00 | $15.75 | |
Claude Sonnet 4.5 | $3.00 | $15.00 | $18.00 | |
Gemini 3 Professional (>200K) | $4.00 | $18.00 | $22.00 | |
Claude Opus 4.6 | $5.00 | $25.00 | $30.00 | |
GPT-5.2 Professional | $21.00 | $168.00 | $189.00 |
What it means for enterprise technical leaders and decision-makers
With the launch of the Qwen3.5 Medium Fashions, the fast iteration and fine-tuning as soon as reserved for well-funded labs is now accessible for on-premise improvement at many non-technical companies, successfully decoupling refined AI from huge capital expenditure.
Throughout the group, this structure transforms how information is dealt with and secured. The power to ingest huge doc repositories or hour-scale movies domestically permits for deep institutional evaluation with out the privateness dangers of third-party APIs.
By working these specialised "Combination-of-Consultants" fashions inside a non-public firewall, organizations can preserve sovereign management over their information whereas using native "pondering" modes and official tool-calling capabilities to construct extra dependable, autonomous brokers.
Early adopters on Hugging Face have particularly lauded the mannequin’s potential to "slender the hole" in agentic eventualities the place beforehand solely the biggest closed fashions might compete.
This shift towards architectural effectivity over uncooked scale ensures that AI integration stays cost-conscious, safe, and agile sufficient to maintain tempo with evolving operational wants.
